gallabytes

• gallabytes 25 Mar 2019 7:47 UTC

  1 point

  in reply to: Charlie Steiner's comment on: How rapidly are GPUs im­prov­ing in price perfor­mance?

  Of course, if you don’t like how an ex­po­nen­tial curve fits the data, you can always change mod­els—in this case, prob­a­bly to a curve with 1 more free pa­ram­e­ter (in­di­cat­ing a de­gree of slow­down of the ex­po­nen­tial growth) or 2 more free pa­ram­e­ters (to have 2 differ­ent ex­po­nen­tials stitched to­gether at a spe­cific point in time).

  Oh that’s ac­tu­ally a pretty good idea. Might redo some anal­y­sis we built on top of this model us­ing that.

• gallabytes 15 Mar 2019 8:30 UTC

  1 point

  in reply to: Benquo's comment on: Black­mailers are pri­va­teers in the war on hypocrisy

  cor­rect. ed­ited to make this more obvious

• gallabytes 14 Mar 2019 11:13 UTC

  15 points

  on: Black­mailers are pri­va­teers in the war on hypocrisy

  This ar­gu­ment would make much more sense in a just world. In­for­ma­tion that should dam­age some­one is very differ­ent from in­for­ma­tion that will dam­age some­one. With black­mail you’re op­ti­mized to max­i­mize dam­age to the tar­get, and I ex­pect tails to mostly come apart here. I don’t see too many cases of black­mail re­plac­ing MeToo. When was the last time the Na­tional En­quirer was a valuable whistle­blower?

  EDIT: fixed some wording

• gallabytes 14 Dec 2018 7:02 UTC

  6 points

  in reply to: Charlie Steiner's comment on: How rapidly are GPUs im­prov­ing in price perfor­mance?

  When try­ing to fit an ex­po­nen­tial curve, don’t weight all the points equally

  We didn’t. We fit a line in log space, but weighted the points by sqrt(y). The rea­son we did that is be­cause it doesn’t ac­tu­ally ap­pear lin­ear in log space.

  This is what it looks like if we don’t weight them. If you want to bite the bul­let of this be­ing a bet­ter fit, we can bet about it.

How rapidly are GPUs im­prov­ing in price perfor­mance?

• gallabytes 18 Nov 2018 8:04 UTC

  3 points

  in reply to: Wei_Dai's comment on: Act of Charity

  I’d op­ti­mize more for not mak­ing en­e­mies or alienat­ing peo­ple than for mak­ing peo­ple re­al­ize how bad the situ­a­tion is or join­ing your cause.

  Why isn’t this a fully gen­eral ar­gu­ment for never rock­ing the boat?

• gallabytes 18 Nov 2018 7:46 UTC

  9 points

  in reply to: jessicata's comment on: Act of Charity

  Based on my mod­els (such as this one), the chance of AGI “by de­fault” in the next 50 years is less than 15%, since the cur­rent rate of progress is not higher than the av­er­age rate since 1945, and if any­thing is lower (the in­sights model linked has a bias to­wards list­ing re­cent in­sights).

  Both this com­ment and my other com­ment are way un­der­stat­ing our be­liefs about AGI. After talk­ing to Jes­sica about it offline to clar­ify our real be­liefs rather than just play­ing games with plau­si­ble de­ni­a­bil­ity, my ac­tual prob­a­bil­ity is be­tween 0.5 and 1% in the next 50 years. Jes­sica can con­firm that hers is pretty similar, but prob­a­bly weighted to­wards 1%.

• gallabytes 18 Nov 2018 1:44 UTC

  11 points

  in reply to: Wei_Dai's comment on: Act of Charity

  I think I’m more skep­ti­cal than you are that it’s pos­si­ble to do much bet­ter (i.e., build func­tional in­for­ma­tion-pro­cess­ing in­sti­tu­tions) be­fore the world changes a lot for other rea­sons (e.g., su­per­in­tel­li­gent AIs are in­vented)

  Where do you think the su­per­in­tel­li­gent AIs will come from? AFAICT it doesn’t make sense to put more than 20% on AGI be­fore mas­sive in­ter­na­tional in­sti­tu­tional col­lapse, even be­ing fairly char­i­ta­ble to both AGI pro­jects and prospec­tive longevity of cur­rent in­sti­tu­tions.

• gallabytes 2 Dec 2017 10:21 UTC

  0 points

  AF

  on: Where does ADT Go Wrong?

  When con­sid­er­ing an em­bed­der $F$, in uni­verse $U$, in re­sponse to which SADT picks policy $\pi$, I would be tempted to ap­ply the fol­low­ing co­her­ence con­di­tion:

  $$E\left[F\right(\pi \left)\right]=E\left[F\right(DDT\left)\right]=E\left[U\right]$$

  (all ap­prox­i­mately of course)

  I’m not sure if this would work though. This is definitely a nec­es­sary con­di­tion for rea­son­able coun­ter­fac­tu­als, but not ob­vi­ously suffi­cient.

• gallabytes 17 Mar 2017 4:23 UTC

  5 points

  in reply to: Alicorn's comment on: I Want To Live In A Baugruppe

  I’m fairly in­ter­ested but don’t re­ally want to be around chil­dren.

Asymp­totic De­ci­sion Theory

• gallabytes 15 Jul 2016 21:07 UTC

  0 points

  AF

  in reply to: orthonormal's comment on: A new pro­posal for log­i­cal counterfactuals

  By cen­sor­ing I mean a spe­cific tech­nique for forc­ing the con­sis­tency of a pos­si­bly in­con­sis­tent set of ax­ioms.

  Sup­pose you have a set of de­duc­tion rules $D$ over a lan­guage $\ell$. You can con­struct a func­tion $f_D:P\left(\ell \right)\to P\left(\ell \right)$ that takes a set of sen­tences $S$ and out­puts all the sen­tences that can be proved in one step us­ing $D$ and the sen­tences in $S$. You can also con­struct a cen­sored $f_D^{\prime }$ by let­ting $f_D^{\prime }\left(S\right)=\{\varphi |\varphi \in f_D(S)\wedge \neg \varphi \notin S\}$.

A new pro­posal for log­i­cal counterfactuals

• gallabytes 17 Dec 2015 21:05 UTC

  2 points

  in reply to: PeteMichaud's comment on: The art of griev­ing well

  I feel like this com­ment be­longs on the LessWrong 2.0 ar­ti­cle (to the point that I as­sumed that’s where it was when I was told about it), but it doesn’t ac­tu­ally mat­ter.

• gallabytes 3 Jun 2015 20:15 UTC

  1 point

  in reply to: 11kilobytes's comment on: Open Thread, Jun. 1 - Jun. 7, 2015

  I’d be in­ter­ested to read an­other take on it if there’s some novel as­pect to the ex­pla­na­tion. Do you have a par­tic­u­lar ap­proach to ex­plain­ing it that you think the world doesn’t have enough of?

• gallabytes 27 Jan 2015 8:14 UTC

  0 points

  on: Meetup : San Fran­cisco Meetup: Pri­soner’s Dilema Tournament

  The tour­na­ment definitely made for an in­ter­est­ing game, but more as a con­ver­sa­tion starter than a game it­self. I sus­pect part of the is­sue was how short this par­tic­u­lar game was, but the con­stant co­op­er­a­tion got old pretty quickly. Some ideas for next time some­one wants to run a dilemma tour­na­ment:

  1) Bracket play—you play with an­other per­son un­til one of you defects while the other co­op­er­ates. Ex­am­ple round: both co­op­er­ate, both co­op­er­ate both defect, both defect, player 1 defects and player 2 co­op­er­ates, player 1 moves on. This has a triv­ial vuln­er­a­bil­ity in that always defect­ing always wins or ties, so there’d have to be some other in­cen­tive to en­courage co­op­er­a­tion.

  2) Fac­tions—each per­son is part of a fac­tion, you have some minor term for the to­tal util­ity of your fac­tion at the end of the game. You could go fur­ther with this and have a nega­tive (even more minor) term for the to­tal util­ity of other fac­tions, or maybe add the op­tion of giv­ing points to other play­ers.

  Un­re­lated note: In the fu­ture there should prob­a­bly be some co­or­di­na­tion around walk­ing to/​from. While noth­ing hap­pened this time, Civic Cen­ter did not feel like the safest place to be walk­ing around alone late at night.

• gallabytes 26 Sep 2014 6:14 UTC

  3 points

  in reply to: KatjaGrace's comment on: Su­per­in­tel­li­gence Read­ing Group 2: Fore­cast­ing AI

  I doubt there is a sharp dis­tinc­tion be­tween them

  Ac­tu­ally, let’s taboo weak and strong AI for a mo­ment.

  By weak AI I mean things like video game AI, self driv­ing cars, WolframAlpha, etc.

  By strong AI I think I mean some­thing that can cre­ate weak AIs to solve prob­lems. Some­thing that does what I mean by this likely in­cludes a gen­eral in­fer­ence en­g­ine. While a self driv­ing car can use its nav­i­ga­tion pro­grams to figure out lots of in­ter­est­ing routes from a to b, if you tell it to go from Cal­ifor­nia to Ja­pan it won’t start build­ing a boat

• gallabytes 24 Sep 2014 0:11 UTC

  4 points

  in reply to: KatjaGrace's comment on: Su­per­in­tel­li­gence Read­ing Group 2: Fore­cast­ing AI

  What’s your rea­son for think­ing weak AI leads to strong AI? Gen­er­ally, weak AI seems to take the form of do­main-spe­cific cre­ations, which provide only very weak gen­eral ab­strac­tions.

  One ex­am­ple that peo­ple pre­vi­ously thought would lead to gen­eral AI was chess play­ing. And sure, the de­sign of chess play­ing AI forced some in­ter­est­ing de­vel­op­ment of effi­cient travers­ing of large search spaces, but as far as I can tell it has only done so in a very weak way, and hasn’t con­tributed mean­ingfully to any­thing re­sem­bling the effi­ciency of hu­man-style chunk­ing.

• gallabytes 23 Sep 2014 23:58 UTC

  3 points

  in reply to: KatjaGrace's comment on: Su­per­in­tel­li­gence Read­ing Group 2: Fore­cast­ing AI

  AGI take­off is an event we as a cul­ture have never seen be­fore, ex­cept in pop­u­lar cul­ture. So, that in mind, re­porters draw on the only good refer­ence points the pop­u­la­tion has, sci fi.

  What would sane AI re­port­ing look like? Is there a way to talk about AI to peo­ple who have only been ex­posed to the cul­tural back­ground (if even that) in a way that doesn’t ei­ther bore them or look at least as bad as this?

• gallabytes 23 Sep 2014 23:53 UTC

  1 point

  in reply to: paulfchristiano's comment on: Su­per­in­tel­li­gence Read­ing Group 2: Fore­cast­ing AI

  Per­haps peo­ple who are a step re­moved from the ac­tual AI re­search pro­cess? When I say that, I’m think­ing of peo­ple like Robin Han­son and Nick Bostrom, whose work de­pends on AI but isn’t ex­plic­itly about it.